google/jigsaw_unintended_bias

Dataset Card for Jigsaw Unintended Bias in Toxicity Classification

Dataset Summary

The Jigsaw Unintended Bias in Toxicity Classification dataset comes from the eponymous Kaggle competition.

Please see the original data description for more information.

Supported Tasks and Leaderboards

The main target for this dataset is toxicity prediction. Several toxicity subtypes are also available, so the dataset can be used for multi-attribute prediction.

See the original leaderboard for reference.

Languages

English

Dataset Structure

Data Instances

A data point consists of an id, a comment, the main target, the other toxicity subtypes as well as identity attributes.

For instance, here's the first train example.

{
    "article_id": 2006,
    "asian": NaN,
    "atheist": NaN,
    "bisexual": NaN,
    "black": NaN,
    "buddhist": NaN,
    "christian": NaN,
    "comment_text": "This is so cool. It's like, 'would you want your mother to read this??' Really great idea, well done!",
    "created_date": "2015-09-29 10:50:41.987077+00",
    "disagree": 0,
    "female": NaN,
    "funny": 0,
    "heterosexual": NaN,
    "hindu": NaN,
    "homosexual_gay_or_lesbian": NaN,
    "identity_annotator_count": 0,
    "identity_attack": 0.0,
    "insult": 0.0,
    "intellectual_or_learning_disability": NaN,
    "jewish": NaN,
    "latino": NaN,
    "likes": 0,
    "male": NaN,
    "muslim": NaN,
    "obscene": 0.0,
    "other_disability": NaN,
    "other_gender": NaN,
    "other_race_or_ethnicity": NaN,
    "other_religion": NaN,
    "other_sexual_orientation": NaN,
    "parent_id": NaN,
    "physical_disability": NaN,
    "psychiatric_or_mental_illness": NaN,
    "publication_id": 2,
    "rating": 0,
    "sad": 0,
    "severe_toxicity": 0.0,
    "sexual_explicit": 0.0,
    "target": 0.0,
    "threat": 0.0,
    "toxicity_annotator_count": 4,
    "transgender": NaN,
    "white": NaN,
    "wow": 0
}

Data Fields

• id: id of the comment
• target: value between 0(non-toxic) and 1(toxic) classifying the comment
• comment_text: the text of the comment
• severe_toxicity: value between 0(non-severe_toxic) and 1(severe_toxic) classifying the comment
• obscene: value between 0(non-obscene) and 1(obscene) classifying the comment
• identity_attack: value between 0(non-identity_hate) or 1(identity_hate) classifying the comment
• insult: value between 0(non-insult) or 1(insult) classifying the comment
• threat: value between 0(non-threat) and 1(threat) classifying the comment
• For a subset of rows, columns containing whether the comment mentions the entities (they may contain NaNs):
  • male
  • female
  • transgender
  • other_gender
  • heterosexual
  • homosexual_gay_or_lesbian
  • bisexual
  • other_sexual_orientation
  • christian
  • jewish
  • muslim
  • hindu
  • buddhist
  • atheist
  • other_religion
  • black
  • white
  • asian
  • latino
  • other_race_or_ethnicity
  • physical_disability
  • intellectual_or_learning_disability
  • psychiatric_or_mental_illness
  • other_disability
• Other metadata related to the source of the comment, such as creation date, publication id, number of likes, number of annotators, etc:
  • created_date
  • publication_id
  • parent_id
  • article_id
  • rating
  • funny
  • wow
  • sad
  • likes
  • disagree
  • sexual_explicit
  • identity_annotator_count
  • toxicity_annotator_count

Data Splits

There are four splits:

• train: The train dataset as released during the competition. Contains labels and identity information for a subset of rows.
• test: The train dataset as released during the competition. Does not contain labels nor identity information.
• test_private_expanded: The private leaderboard test set, including toxicity labels and subgroups. The competition target was a binarized version of the toxicity column, which can be easily reconstructed using a >=0.5 threshold.
• test_public_expanded: The public leaderboard test set, including toxicity labels and subgroups. The competition target was a binarized version of the toxicity column, which can be easily reconstructed using a >=0.5 threshold.

Dataset Creation

Curation Rationale

The dataset was created to help in efforts to identify and curb instances of toxicity online.

Source Data

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

[More Information Needed]

Annotations

Annotation process

[More Information Needed]

Who are the annotators?

[More Information Needed]

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

This dataset is released under CC0, as is the underlying comment text.

Citation Information

No citation is available for this dataset, though you may link to the kaggle competition

Contributions

Thanks to @iwontbecreative for adding this dataset.